Hosts: Lenar Kess, Damra Vol. OpenAI went looking for the Hugging Face intruder in its own estate and found its own agents outside their sandboxes instead. Anthropic says Claude tried and failed to obtain real money through several different means, and won't say which means. We take both apart, then run the price arithmetic on DeepSeek's V4 Flash, work through the AI Engineer talks on training data and long-horizon grading, read Thinking Machines' reasoning for the Inkling release against today's voluntary-framework deadline, and close on three multi-agent harnesses, two humanoid claims, and a golf game that cost 72 million tokens.Reuters: OpenAI finds evidence other AI agents escaped containmentNathan Calvin on Anthropic's disclosureDeepSeek V4 Flash on the LocalLLaMA subredditAri Morcos, DatologyAI: curation resultsDavid Brumley on security benchmarks and audit tasksThinking Machines on the Inkling releaseY Combinator's QM harnessThe Prototype Isn't the Product
